This paper deals with the early detection of small parameter variations in non-linear systems. The objective is to detect these variations, before they manifest themselves as a sudden change in the phase behavior. It is assumed that these parameter variations occur slowly in time. Early detection of slow time-scale anomalous behavior in nonlinear dynamical system is achieved by observing and inferring time series data of the state or output variables, at the fast time-scale. The detection algorithm makes use of Symbolic Dynamics and Finite State Automata. As an example, we consider the Duffing equation which is a second order non-linear differential equation. The methodology is validated both in simulation and experiment. This concept can be further extended to dynamical systems of higher complexity.